KASALv2: Fully Automatic 3D Rotational Symmetry Classification and Axis Localization

Conference: CVPR 2026
Paper: CVF Open Access
Code: https://github.com/WangYuLin-SEU/KASAL
Area: 3D Vision
Keywords: Rotational Symmetry, 6D Pose Estimation, Axis Localization, Geometric Prior, Reference-free Analysis

TL;DR

KASALv2 proposes a fully automatic framework that identifies rotational symmetry types, rotation orders, and all canonical axes of 3D objects at once without any reference geometry. It covers all 8 canonical rotational symmetry types, achieving 94.75% accuracy on 438 symmetric objects from GSO. Feeding the estimated symmetry priors into FoundationPose training improves pose estimation accuracy by up to 0.9% across 5 BOP datasets.

Background & Motivation

Background: Rotational symmetry is a crucial geometric prior in 6D pose estimation—it resolves pose ambiguity, stabilizes optimization, and supports "symmetry-aware" evaluation metrics such as ADD-S, MSSD, and MSPD. Recent systems like HccePose(BF) and ZebraPose use symmetry priors to map canonical poses to their nearest equivalent representation to improve precision.

Limitations of Prior Work: Obtaining symmetry annotations currently relies almost entirely on manual or semi-automatic methods, which often require pre-specifying the symmetry type or order. Manual annotation is impractical for datasets with tens of thousands of 3D models (e.g., large-scale synthetic training for unseen objects, mesh generation, or 3D asset processing). Existing automatic methods either cover only partial symmetry types or depend on geometric references like centroids/keypoints, making them sensitive to noise and occlusion. Wang et al., although closest to full coverage, still requires manual specification of type and order.

Key Challenge: 3D rotational symmetry structures are highly diverse—corresponding to orientation-preserving subgroups of SO(3), including continuous types \(C_\infty/D_\infty\), finite cyclic groups \(C_n\), dihedral groups \(D_n\), and rotational groups of Platonic solids \(A_4/S_4/A_5\), totaling 8 canonical types. Each has different axis counts, orders, and inter-axis relations. Achieving "fully automatic + full coverage" necessitates a unified framework that jointly reasons about type and axis structure rather than applying per-type rules.

Goal: Upgrade the task from "locating axes given a type" to "automatically determining symmetry type + rotation order + all canonical axes without types or references."

Key Insight: The authors observe that high-order symmetry axes naturally satisfy the periodicity of low-order rotations. Consequently, they form deeper "descent basins" in alignment loss—optimization is automatically attracted toward dominant high-order axes without needing a prior order.

Core Idea: Organize the 8 symmetry types into a top-down "Geometric Degeneration Hierarchy" (GDH). First, lock dominant high-order axes and estimate orders in a reference-free manner using self-consistency. Then, use inter-axis tilt angles constrained by the GDH to verify the type and reconstruct all canonical axes. Finally, handle order degeneration caused by appearance through texture expansion.

Method

Overall Architecture

The input is a 3D object (point cloud/mesh) with an unknown symmetry type, and the output is its symmetry type, rotation order, and a full set of canonical axes. The "theoretical foundation" is the Geometric Degeneration Hierarchy (GDH): it reorganizes all SO(3) rotational subgroups into a computable hierarchical tree based on how the "number of axes/order decreases with geometric degeneration," providing constraints on feasible types and inter-axis tilts. Guided by GDH, the system follows: dominant high-order axis localization → order estimation → candidate type narrowing and verification via secondary axis tilt → full canonical axis reconstruction → texture symmetry expansion. The entire pipeline is reference-free and requires no human input.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: 3D object with unknown type"] --> B["Geometric Degeneration Hierarchy (GDH)<br/>Organizes 8 symmetry types into<br/>computable degeneration constraints"]
    B --> C["Dominant axis localization + Order recognition<br/>Mixed-order alignment loss + Two-stage sampling<br/>Periodic self-similarity estimates N"]
    C --> D["Hierarchical classification & Full axis reconstruction<br/>Verify type via secondary axis tilt αF<br/>Recover all canonical axes via template alignment"]
    D --> E["Texture symmetry expansion<br/>Appearance reduces geometric order to its factors"]
    E --> F["Output: Symmetry type + Order + Full canonical axes"]

Key Designs

1. Geometric Degeneration Hierarchy (GDH): Turning 8 symmetry types into a computable tree

Existing "all-type" methods require manual type specification because they lack a structured prior for unified reasoning. GDH reorganizes SO(3) orientation-preserving subgroups into a hierarchy: the top is the isotropic sphere (any direction is an axis), branching into two degeneration lineages—Dihedral-Cyclic and Platonic. The Dihedral-Cyclic lineage starts from cylindrical \(D_\infty\) and degenerates via two paths: discretization of continuous rotation (\(D_\infty\Rightarrow D_n, C_\infty\Rightarrow C_n\) for \(n\ge2\)) and loss of orthogonal 2-fold axes (\(D_\infty\Rightarrow C_\infty, D_n\Rightarrow C_n\)). The Platonic lineage describes discrete degeneration of multi-high-order axis structures (\(A_5\Rightarrow A_4, S_4\Rightarrow A_4\)). The key to GDH is that it encodes "how axis counts, orders, and inter-axis relations evolve with degeneration" into computable constraints. Locating a dominant axis narrows sub-classes; estimated orders further compress the candidate set; finally, verifying a secondary axis satisfying the required tilt locks the type.

2. Dominant high-order axis localization + Order recognition: "Automatically sucking out" high-order axes via periodic self-consistency

Locating axes without reference is difficult because alignment error cannot be defined without knowing the order. The authors use a mixed-order alignment loss to aggregate alignment errors across multiple candidate orders: \(L_\text{mix}(a)=\sum_{n\in N}L_n(a)\), where \(L_n(a)=\sum_{k=1}^{n-1}\text{Chamfer}(P,\,R(a,\theta_k)P)\) and \(\theta_k=2\pi k/n\). Since high-order axes satisfy low-order periodicity, they yield lower aggregated loss and form deeper basins, naturally biasing optimization toward dominant high-order directions (candidate set \(N=\{3,4,5,6\}\)). Localization is two-stage: coarse uniform sampling on the upper hemisphere (128 directions) to find top-\(k_\text{cand}\) low-loss regions, followed by fine sampling within a \(10^\circ\) spherical cap and Adam optimization (lr=0.04, 5 steps). After locking the dominant axis \(a^*\), a rotation self-similarity signal is generated by calculating Chamfer distances over dense rotations (typically 360 angles). The main frequency gives a coarse estimate \(N_\text{fft}\), and \(N_\text{est}\) is selected from \(\{N_\text{fft}, N_\text{fft}\pm1\}\) based on minimum reconstruction error. Ambiguous cases are verified hierarchically (\(N_\text{est}\ge3\) is discrete rotational; \(=2\) uses a \(180^\circ\) test to distinguish 2-fold from reflection; \(\le1\) uses a \(45^\circ\) test to distinguish continuous from asymmetric).

3. Hierarchical type classification and full axis reconstruction: Finalizing via tilt angle \(\alpha_F\) and template alignment

With the main axis and order, GDH narrows the hypothesis space to subsets compatible with \(N_\text{est}\). Remaining ambiguity lies in the orientation of secondary axes relative to the main axis. Each candidate type in GDH specifies a characteristic inter-axis tilt \(\alpha_F\) (invariant to pose and scale). A secondary axis is searched along a circular trajectory centered at \(a^*\) with radius \(\alpha_F\): \(b^*=\arg\min_{b\in S(\alpha_F)}L_\text{Chamfer}(b;N)\), followed by 1D gradient refinement of top-k seeds. The type is determined by the order pair \((N_{a^*},N_{b^*})\) and the measured tilt: continuous \(C_\infty/D_\infty\) checks for discrete periodicity or orthogonal 2-fold axes; discrete \(C_n/D_n\) checks for orthogonal 2-fold axes; Platonic types are identified by comparing measured tilts to theoretical constants: \(70.53^\circ\) (\(A_4\)), \(90^\circ\) (\(S_4\)), and \(63.43^\circ\) (\(A_5\)). Once the type is fixed, GDH provides a canonical axis template. A rigid transformation \(Q\) is estimated to align the template pair \((\hat a,\hat b)\) to the detected pair \((a^*,b^*)\), and full axes are recovered as \(u_i^\text{final}=Q\,\hat u_i\).

4. Texture symmetry expansion: Modeling appearance as second-order geometric degeneration

Real object geometry is rarely perfectly regular, and surface texture/color can break rotational invariance in the image domain. The authors treat texture as secondary order degeneration superimposed on geometric symmetry: it only changes the effective order of axes, not their orientation, and the effective order must be a factor of the geometric order—\(n_\text{tex}\in\{d\mid d\text{ divides }n_\text{geo}\}\cup\{1\}\). Thus, a 6-fold geometric object might appear 3-fold or 2-fold under periodic texturing. Given the recovered geometric type and axis configuration, texture refinement is calculated by analytically evaluating appearance consistency around each axis.

Loss & Training

The method is training-free geometric optimization: all alignment losses are based on Chamfer distance, with Adam used for few-step refinement (main axis lr=0.04, 5 steps; secondary axis lr=0.05, 5 steps). The sampling budget is driven by a single parameter, the spherical cap radius \(r\). The coarse sampling lower bound is derived from the area ratio \(N_\text{min}=\frac{2}{1-\cos r}\) (limited to the upper hemisphere, rounded to multiples of 8). The experiments ran on i5-13400F + RTX 3050 + 32GB, taking ~1.3–1.8s per object. The only "training" involved is downstream: FoundationPose is trained on ~1M rendered images using KASALv2-generated priors injected via ZebraPose-style symmetry-aware normalization.

Key Experimental Results

Main Results

Evaluation across DSRSTO, BOP, and GSO datasets using type accuracy \(acc_t\), order accuracy \(acc_o\), total accuracy \(acc_T\), normalized axis alignment error \(e_{ADI}/d\) (\(d\) is object diameter), and runtime.

Dataset	Scale	\(acc_T\)	\(e_{ADI}/d\) (Ours / Wang et al.)	Time (s)
GSO (438 symmetric)	438 of 944 scans	94.75%	0.00262 / 0.00256	1.37
BOP (50 re-annotated)	Real-world objects	84.00% (80.00% orig)	0.00256 / 0.00290	1.61
DSRSTO	38 CAD, all types	81.48%	0.00212 / 0.00172	1.46
DSRSTO\(_\text{tex}\) (texture)	11 texture models	72.32%	0.00195 / 0.00181	1.16

On GSO, \(acc_t\) reached 96.58% and \(acc_o\) 98.10%. Following BOP re-annotation, \(e_{ADI}/d\) dropped from 0.00290 to 0.00256, indicating localization precision superior to Wang et al. \(C_n\) remains the lowest accuracy category (e.g., 83.53% on GSO).

Ablation Study

Investigation of the candidate order set \(N\) on DSRSTO (optimal \(r=10^\circ\) gives \(acc_T=81.48\%\)):

Order Set \(N\)	\(acc_T\)	\(e_{ADI}/d\)	Time (s)	Description
{3,4}	70.37%	0.00419	1.02	Insufficient set
{3,4,5}	77.78%	0.00423	1.20	Still lacking
{3,5,6}	77.78%	0.00411	1.69	Misclassifies Cubic/Dihedral without 4-fold
{3,4,5,6}	81.48%	0.00418	1.46	Minimum effective config (Default)
{3,4,5,6,7}	81.48%	0.00419	1.77	Expansion yields no gain
{2,3,4,5,6}	77.78%	0.00659	2.28	2-fold introduces instability

Key Findings

• {3,4,5,6} is the minimum effective configuration: Removing any key order (especially 4-fold) makes it difficult to distinguish Platonic from Dihedral symmetry; further expansion to 7 or 9 only increases time.
• 2-fold is harmful: Trivial half-turn alignment dominates optimization, misclassifying 3-fold high-order models as 2-fold, worsening \(e_{ADI}/d\) from 0.00418 to 0.00659. This justifies starting the set from 3.
• Downstream Pose Improvement: FoundationPose trained with KASALv2 priors improved by up to 0.9% on 5 BOP datasets, proving the real-world value of automatically estimated rotation priors.

Highlights & Insights

• "High-order axes naturally fall into deeper loss basins": This is the pivot of the paper. Because high-order rotations satisfy low-order periodicity, the mixed-order loss can "suck out" the dominant axis without a prior order, elegantly bypassing the chicken-and-egg problem.
• GDH turns classification into computable constraints: Unlike works that merely list the 8 types, KASALv2 encodes branch relations into step-by-step constraints (Main Axis → Order → Secondary Tilt), making fully automatic reasoning operational.
• Factor-based texture modeling: Formalizing "appearance reduces 6-fold symmetry to 3-fold/2-fold" as "effective order must divide geometric order" is a clean approach that preserves axis orientation while avoiding re-optimization for textures.
• Near-zero deployment cost: No training and ~1–2s per object on an RTX 3050 makes batch-processing symmetry labels for tens of thousands of models feasible.

Limitations & Future Work

• \(C_n\) category weakness: Accuracy for \(C_n\) remains significantly lower across datasets, attributed to difficult distinctions between high-order and visual ambiguities.
• Dependence on geometric regularity: The core is Chamfer self-alignment; robustness to scan noise, incomplete point clouds, or severe occlusion was not fully stress-tested in the paper.
• Idealized texture expansion: Modeling texture strictly as factor-based periodicity might not apply to non-periodic/local textures or pseudo-symmetries caused by lighting.
• Modest downstream gain: A 0.9% improvement is limited; exploring more powerful symmetry-aware training methods beyond simple label normalization might amplify the benefits.

Related Work & Insights

• vs Wang et al. (Key-axis localization): Direct extension. Wang handles all types but requires manual type/order specification and uses exhaustive direction enumeration. KASALv2 automates this via GDH constraints and outperforms it on BOP (\(e_{ADI}/d\) 0.00256 vs 0.00290).
• vs Reference-based methods (Centroids/PCA): Reference-based methods are sensitive to surface noise and occlusion; KASALv2 is reference-free and relies on global spherical alignment.
• vs Limited-coverage reference-free methods (e.g., Hruda et al., Rajkumar et al.): Prior works often fail on continuous symmetry (cylinders/cones) or specific discrete types. KASALv2 covers all 8 canonical types in one framework.
• vs Pose-end symmetry utilization (HccePose, ZebraPose): These use pre-defined symmetry for pose normalization; KASALv2 provides the automatic upstream annotation to enable this.

Rating

• Novelty: ⭐⭐⭐⭐⭐ GDH + high-order self-consistent localization enables fully automatic analysis for the first time.
• Experimental Thoroughness: ⭐⭐⭐⭐ Three datasets + dual ablations + downstream verification, though \(C_n\) failure analysis is relatively brief.
• Writing Quality: ⭐⭐⭐⭐ Clear progression from theory to operation; symbols are well-standardized.
• Value: ⭐⭐⭐⭐ Transforms symmetry annotation from a manual bottleneck into a batch-automated process for 6D pose and 3D asset pipelines.

Related Papers

• [CVPR 2026] Fusion of Depth and Semantics for Probabilistic Floorplan Localization
• [CVPR 2026] CoLoR: The Devil is in Scene Coordinate Regression for Large-Scale Visual Localization
• [CVPR 2026] Long-SCOPE: Fully Sparse Long-Range Cooperative 3D Perception
• [CVPR 2026] 3D-IDE: 3D Implicit Depth Emergent
• [CVPR 2026] ActionMesh: Animated 3D Mesh Generation with Temporal 3D Diffusion